We propose an adversarial data augmentation technique (top figure) for increasing shape sensitivity in vision models.  Across model architectures & augmentations, shape sensitivity correlates with increased robustness of learned models (bottom figures),  With ELeaS, we improve both metrics over and above SOTA augmentation (gold diamond (E), vs blue triangle, (T)). 

Edges to shapes to concepts: Adversarial augmentation for robust vision

A. Tripathi, R. Singh, A. Chakraborty, P. Shenoy. CVPR 2023.

We propose a coreset selection approach for managing the replay buffer in replay-based continual learning. Our optimization-driven approach significantly improves continual learning metrics--upto 10%/ 18% relative accuracy and forgetting--on a range of datasets and CL settings (class- & task-incremental, online & offline), compared to SOTA CL approaches.

GCR: Gradient coreset based replay buffer selection for continual learning

R. Tiwari, K. Killamsetty, R. Iyer, P. Shenoy. CVPR 2022.

We propose a novel image-to-image learning paradigm as a testbed for end-to-end learning of neurosymbolic architectures. We design & evaluate NSN-net, a learning architecture with a an algorithmic core and encoders-decoders for interacting with input-output supervision. Our prototype accurately solves visual sudoku / visual maze with no intermediate supervision. 

End-to-End Neuro-Symbolic Architecture for Image-to-Image Reasoning Tasks

A. Agarwal, P. Shenoy, Mausam. Arxiv 2021.